Abstract
Performance evaluation of urban autonomous vehicles (AVs) requires a realistic model of the behavior of other road users in the environment. Learning such models from data involves collecting naturalistic data of real-world human behavior. In many cases, acquisition of this data can be prohibitively expensive or intrusive. Additionally, the available data often contain only typical behaviors and exclude behaviors that are classified as rare events. To evaluate the performance of AVs in such situations, we develop a model of traffic behavior based on the theory of bounded rationality. Based on experiments performed on large naturalistic driving data, we show that the developed model can be applied to estimate the probability of rare events, as well as to generate new traffic situations.
Introduction
To guarantee the safety of Autonomous Vehicles (AVs), field testing alone is statistically insufficient—an AV would need to drive billions of miles to prove its reliability. Consequently, the industry relies heavily on simulation. However, for simulations to be useful, the simulated human drivers must behave realistically.
The core problem is that real-world naturalistic driving datasets mostly contain typical, safe behaviors. "Rare events" like crashes and near-misses are, by definition, missing or extremely sparse. Existing rare event (RE) sampling methods, like Cross-Entropy (CE) importance sampling, mathematically force simulations into failure states to test the AV. However, these methods often lack interpretability; they tweak statistical distributions to cause a crash, but they do not explain why the human driver caused the crash in terms of actual driving behavior.
Methodology: Bounded Rationality in Traffic
The authors tackle this by applying the theory of bounded rationality—a concept from behavioral economics and AI which posits that agents (human drivers) do not always make mathematically optimal decisions. Instead of just perturbing physics or trajectories, the authors model human drivers using a set of utility functions based on safety (distance gap, time-to-collision) and progress (velocity).
They introduce a rationality vector ($\Lambda$) that dictates how strictly a driver adheres to these utilities. The methodology involves three main steps:
- Behavior Categorization: By varying the rationality vector, the authors categorize drivers into 8 distinct behavior profiles for a "cut-in" scenario (e.g., aggressive high-speed cut-ins vs. cautious low-speed cut-ins).
- Parameter Optimization: They use a Simulated Annealing (SA) heuristic to find the specific behavior category and rationality parameters that maximize the likelihood of a rare event (defined here as a near-crash where the distance gap drops below 0.01 meters).
- Situation Generation: They extend the model with mixing parameters to blend different behavior categories, allowing the model to generate synthetic data that mimics the multimodal nature of real-world human driving.
The study utilized the University of Michigan SPMD (Safety Pilot Model Deployment) dataset, a massive naturalistic driving dataset collected over two years. From this, the authors extracted 74,449 vehicle cut-in events, capturing trajectories for 5 seconds immediately following the maneuver.
Results and Analysis
The Bounded Rationality (BR) approach demonstrates significant advantages over traditional methods in terms of speed, accuracy, and interpretability.
Efficiency and Variance Reduction
The BR approach is $10^4$ times faster than Crude Monte Carlo (CMC) sampling and provides a 33% speedup over standard Cross-Entropy (CE) sampling. Crucially, the BR method achieved a 99% reduction in variance compared to CE. While both CE and BR successfully force rare events at high rates ($\sim3-4 \times 10^{-2}$ compared to CMC's $5.6 \times 10^{-4}$), BR's lower variance means its mathematical estimate of the true probability of these rare events is far more accurate.
Behavioral Interpretability and Fidelity
Unlike standard CE methods that operate purely on statistical distributions, the BR method tweaks human motivations. This makes the generated edge cases highly interpretable; for instance, an engineer can see that a simulated crash happened specifically because the simulated driver prioritized speed over time-to-collision. Explicitly modeling the sub-optimality of human decisions is shown to be crucial for realistic safety verification.
Furthermore, the BR model showed high fidelity to real data. When generating synthetic driving situations, QQ plots showed a strong linear relation with real-world naturalistic driving data, with Pearson correlation coefficients of $\rho = 0.919$ for time-to-collision and $\rho = 0.98$ for distance gap.
Conclusion
While highly effective, the current approach is primarily demonstrated on a single specific maneuver: two-lane vehicle cut-ins. The utility functions are manually defined for this specific scenario. Future work involves expanding the bounded rationality framework to a wider variety of complex urban situations—such as multi-agent intersection negotiations or interactions with pedestrians—where defining utility functions and behavior categories will be significantly more complex.
Summary updated